Music Tokenization Beats 34x Model Scaling

Changing a music model's note encoding cuts its errors in half, while scaling it 34-fold barely helps, per a new arXiv paper. For teams building generative music tools, that flips the usual playbook. The format decision made before training mattered more than the size of the model trained on it.

The paper, arXiv 2608.03999, held a Qwen3.5 backbone fixed at sizes from 0.8 billion to 27 billion parameters. It also fixed training data, compute budget, and decoding settings, then tested seven different ways to represent musical notes as tokens.

The researchers' own scheme, called PMT, uses 10-millisecond timing resolution, per-note velocity, and separate tracks for texture, built from a 609-symbol vocabulary. At 0.8 billion parameters, PMT scored a Frechet Music Distance (FMD) of 159, the paper's measure of how far generated music strays from real recordings. Beat-grid tokenizations scored 272 to 286, worse even at the full 27-billion-parameter scale.

Anyone still judging a generative model mainly by parameter count is measuring the wrong variable here. The paper doesn't say whether PMT's gains carry over outside music generation, or what its larger vocabulary costs at inference. But on this benchmark, representation beat scale by a wide margin, and it wasn't close.

The result lands as most generative-music work still defaults to scaling the backbone first and testing representation second.